fix: coerce low combined_risk_level to medium instead of rejecting

[asteier2026]

Summary

• The LLM occasionally outputs combined_risk_level='low' with a non-leave_as_is
  protection_method_suggestion (e.g. suppress_inference, generalize)
• Previously EntityDispositionSchema._validate_protection_consistency raised a ValidationError, which
  caused the entire record to be skipped (UPDATE - sometimes the record isn't skipped, but the entity silently gets dropped from the sensitivity disposition)
• Now combined_risk_level is promoted to medium — preserving the model's intent to protect the entity
  rather than suppressing it. This is more appropriate for weaker models that apply protection methods
  without fully accounting for contextual risk
• Updated the corresponding test from asserting a ValidationError to asserting that
  combined_risk_level is promoted and the original protection_method_suggestion is retained

UPDATE - the sensitivity disposition prompt is overloaded and doesn't do as well on gpt-oss-120b. We will do another PR to fix that, but above is a quick fix for now

[greptile-apps]

Greptile Summary

This PR changes how EntityDispositionSchema handles the inconsistent case where an LLM returns combined_risk_level='low' alongside a protective (non-leave_as_is) method. Instead of raising a ValidationError (which could silently drop entities from the sensitivity disposition), the validator now logs a warning and promotes combined_risk_level to medium, preserving the protection intent.

• rewrite.py: The validator coercion is safe — CombinedRiskLevel is a str, Enum, so the in-validator assignment of CombinedRiskLevel.medium is equivalent to the string "medium" throughout serialization and downstream comparisons; use_enum_values=True poses no issue here.
• test_schemas.py: The parametrized test now covers all four active protection methods (replace, generalize, remove, suppress_inference) and asserts both that the risk level is promoted and that the original method is retained. All three previous reviewer concerns (logging, test coverage, and description–code alignment) have been addressed.

Confidence Score: 5/5

Safe to merge — the coercion replaces a silent entity-drop with an explicit risk-level promotion, the warning log is in place, and the tests cover all four protection methods.

The change is narrowly scoped to one validator branch. The coercion is semantically sound, the warning gives operators an observable signal, and the parametrized test catches regressions across every relevant protection method. No downstream consumers of combined_risk_level are broken by the promotion from low to medium.

Files Needing Attention: No files require special attention.

Important Files Changed

Filename	Overview
src/anonymizer/engine/schemas/rewrite.py	Replaces the ValidationError raised for low-risk + non-leave_as_is combos with a warning log and coercion of combined_risk_level to medium; logic is correct and consistent with model semantics.
tests/engine/test_schemas.py	Old 'expects ValidationError' test replaced with a parametrized test covering all four coercible protection methods; both the promoted risk level and the retained protection method are asserted.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["LLM output: combined_risk_level='low'\nprotection_method_suggestion=X"] --> B{"X == leave_as_is?"}
    B -- Yes --> C["Valid: no coercion needed"]
    B -- No --> D["Log warning\n(Entity id, label, method)"]
    D --> E["Promote combined_risk_level → 'medium'\nKeep protection_method_suggestion=X"]
    E --> F{"combined_risk_level == 'high'\nAND method == 'leave_as_is'?"}
    C --> F
    F -- Yes --> G["Raise ValidationError"]
    F -- No --> H["Return validated EntityDispositionSchema\n(protection intent preserved)"]

Loading

_{Reviews (5): Last reviewed commit: "fix: redact entity_value from coercion w..." | Re-trigger Greptile}

[lipikaramaswamy]

One non-blocking serialization consistency suggestion.

[lipikaramaswamy]

Non-blocking: because model_config uses use_enum_values=True, normally validated combined_risk_level values are stored and returned by model_dump() as plain strings. This assignment happens after validation, so the coerced path instead stores a CombinedRiskLevel enum object (although it compares equal to "medium"). Could we assign CombinedRiskLevel.medium.value here and add a model_dump() assertion so coerced and non-coerced records preserve the same serialization contract?

[asteier2026]

Made suggested changs.

[asteier2026]

Note that this branch has a different version of rewrite_generation.py than what is being updated in PR 208 - it doesn't have the prereplace logic. How do we handle that?